Seoyul Oh

A satellite image contains much more information than a specific question needs. RIFT identifies tiles and visual tokens that can be removed without changing the answer. Its two stages—query-conditioned tile pruning and elastic prefill—reduce the energy and latency of vision-language model inference. We evaluate the approach on satellite imagery using LLaVA-1.5-7B and a Jetson AGX Orin.

How traffic arrives at a cellular base station affects how much energy it uses. EcoCell reshapes traffic to exploit energy-saving opportunities without changing cellular protocols or hardware. It can run in the radio access network scheduler or as an independent middlebox. Across five network applications on a production base station, EcoCell saves up to 32% of dynamic energy without reducing quality of experience.

Satellites constantly move across the globe, making separate, region-specific constellations costly and wasteful. MP-LEO proposes decentralized constellations in which multiple parties contribute satellites and share spare capacity. We explore how this design could reduce costs and orbital occupancy, remain robust when participants withdraw, and what technical questions must be addressed to make it work.